Best AI agent answered 39% of hard live-web research tasks in a new benchmark

Researchers from HKUST (Guangzhou) and other universities built OpenDS-World: 200 hand-checked questions that start from a video and need evidence from several live websites, file types and interactive pages. Of 12 models tested with the same tools and a limit of one hour and 100 tool calls per task, GPT-6 Astra did best at 39.0%, followed by Claude Opus 5.5 at 38.0% and Gemini 3.8 Flash at 36.5%. Human testers with a normal browser and no AI got 21.1% across 1,000 attempts, taking about 39 minutes per attempt on average. More tokens or tool calls did not mean better answers, and failed tool calls took 19% to 62% of tool time depending on the model. It is a research benchmark, and answers were graded by another model (Qwen3.8-27B) against reference answers. On the Hugging Face paper page there are only two comments: the submitter restating the title and an automated list of similar papers. No real discussion yet, so no camps and no consensus. Sorted take: deep research agents are useful for finding sources, but on hard multi-step questions even the best one was wrong more often than right here. Treat the answer as a lead, not a finding.